A systematic evaluation of transformer-based architectures for fake news detection in the low-resource Burmese language

The rapid spread of fake information on social media, especially in low-resource languages where standard benchmarks and annotated datasets are scarce, remains a major challenge. Although Burmese is the official language of Myanmar, it has received limited attention in fake news detection research. This study presents a systematic analysis of transformer-based fake news detection systems at the headline level for low-resource Burmese. A balanced sample of 10,000 Burmese news headlines was created to facilitate this analysis under controlled experimental conditions and minimize bias in the data. This study explores multiple paradigms of transformer learning, such as multilingual models with cross-lingual transfer of knowledge, language-specific models trained to capture Burmese linguistic characteristics, and hybrid models trained to combine contextual representations with local feature extraction. The models were compared based on standard classification metrics, cross-validation, and statistical significance analysis. The experimental results demonstrate the highly consistent performance of all transformer-based models, with an accuracy of 0.973 to 0.979, F1-scores of up to 0.979, and ROC-AUC of up to 0.995. The highest overall accuracy was obtained with the language-specific model, whereas the hybrid model was the most stable, with the highest cross-validation mean accuracy (0.988 ± 0.0037). Further analysis indicates that the models perform better on synthetically generated fake news than on socially collected data. It also reflects the influence of linguistic variation in real-world information. Although small performance variations exist, the statistical analysis indicates that the variations are not significant ( p > 0.05), which means that the architectures exhibit statistically comparable performances. The results indicate that dataset quality, controlled construction, and representation learning are more important in low-resource settings than architectural complexity. This study provides benchmark framework for Burmese fake news detection and insights into the model’s behavior, robustness, and generalization under limited data conditions.